Claude Fable is relentlessly proactive(simonwillison.net)
504 points by lumpa 10 hours ago | 403 comments
tl;dr: Claude Fable 5 debugged a CSS scrollbar bug by autonomously inventing elaborate workarounds: spinning up Playwright browsers, writing a PyObjC script to capture Safari window IDs for screenshots, injecting JavaScript into app templates to simulate keystrokes, and building a custom CORS server to exfiltrate DOM measurements from a Web Component's shadow DOM. The session burned ~$12 in tokens to fix a two-line CSS issue, highlighting both the model's impressive proactivity and the serious risk of running such agents outside a sandbox where prompt injection could cause significant damage.
HN Discussion:
  • Fable wastes tokens overengineering simple tasks, making it worse despite being smarter
  • Running coding agents outside a sandbox is reckless yet widely done
  • Calling the model 'proactive' anthropomorphizes token-greedy looping behavior
  • Fable's autonomous problem-solving has been genuinely useful in personal experience
  • ~The behavior likely reflects harness design choices rather than model intent